[Comp-neuro] (2) Levels

James Schwaber schwaber at mail.dbi.tju.edu
Thu Aug 21 23:38:23 CEST 2008


_(2) Levels _

I like what several post's have said about the legitimacy of modeling at 
different levels of scale/analysis. In the first place, despite all the 
talk about multi-scale modeling, actual multi-scale is really hard and 
rare. So working at level 'x' and claiming you really are laying the 
groundwork for level 'y' (in the fullness of time!) can be disingenuous.

Denis Noble makes a wonderful point about the requirement in model 
building at any level of analysis for top down causation to constrain 
bottom up mechanisms. One of his favorite examples is that he would 
never find pacemaker activity at the heart no matter how long he studied 
cardiac myocyte biophysics. Sydney Brenner has called this 'middle out' 
analysis reaching up for the functional behavior and down for the 
mechanistic components.

In contradiction to these folks I understand that you Jim believe if you 
constrain a (model) system from top-down, then that is what you will 
surely find... that the system exhibits the phenomenon you 
told/constrained it to do and that this result is meaningless. They (and 
I from gene network experience) would not agree with you that this is 
easy and automatic as you seem to believe, e.g. the claim that complex 
models can "fit anything" sounds good but is not true. No matter, the 
claims here are (a) the outcome you lightly dismiss is a great hoped for 
outcome/starting point and (b) a model of a functional process can not 
be developed from working at the level of the reduced data alone, it 
requires the function to be explained or understood. The implication, 
which is going to irritate you, is that the function is not inherent 
(unique) in the collection of the parts. The "E word". Say no more.

A more general idea that I find useful (and that I think is common in 
the SBML-ish world I mostly work in) is that models are best considered 
as models of process, not of static structure/function. Even more 
crucial, we evaluate our models as 'a model of what', are they heuristic 
or useful (testable predictions) in the study of that - */_not at all 
meant to be real_/*, or the real object. We see useful models of the 
same larger system of many kinds, some as abstract as graphical patterns.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.neuroinf.org/pipermail/comp-neuro/attachments/20080821/657c81e4/attachment.html


More information about the Comp-neuro mailing list